Adaptive convolutional neural network for detecting safety violations
Abstract
Adaptive convolutional neural network for detecting safety violations
Incoming article date: 23.04.2025The article presents an adaptive convolutional neural network for automated detection of safety violations in real time. Unlike existing solutions using static models, the proposed approach includes two key innovations. Automatic adaptation of model weights with a combination of stochastic and gradient descent methods. The algorithm dynamically adjusts the learning rate and the depth of parameter modification, which makes it possible to preserve previously acquired knowledge while further training on new data without degrading accuracy. Optimized context processing mechanism – the model analyzes not only objects (for example, the absence of a helmet), but also their relative location (a worker in a dangerous area without personal protective equipment), which reduces the number of false alarms. The developed system integrates computer vision, alert generation, and analytics modules, providing not only instant response to violations, but also long-term risk analysis. Experiments have confirmed a 15% increase in accuracy when working in changing lighting conditions and shooting angles.
Keywords: convolutional neural network, information system, industrial accidents, safety, production, model training, neural network, adaptive algorithm